The AI Bubble of 2023
thereformedbroker.com
thereformedbroker.com
This is extremely valuable and saved me hours/was the difference between me stopping or not.
There will surely be people that overhype this, but there is real value as well.
[1] https://stackoverflow.com/questions/58552645/what-exactly-is...
[2] https://stackoverflow.com/questions/57328953/trying-to-decod...
I fully expect a node.js server in a docker container to query GPT for error messages found in logs. They'll give it a cutesy name like ChattyDev, and before you know it, every coding school will be pumping out coders who think it's an essential development tool, and HN threads filled with people who will defend the practice to the death, referring to examples such as left-pad for prior art.
"yeah, well...people made fun of left-pad, but we see how it has completely automated the padding of strings, removing one more chance for error!"
> Wall Street loves to play “Who wins, who loses” when a new technology explodes onto the scene. The fingers were all pointing at Google as a potential loser.
Augment search, yes. Replace it with the current chatGPT interface? No way.
The dotcom bubble was a bubble because there were a lot of junk, overvalued companies. Doesn't mean websites aren't useful.
When the bubble pop as it inevitably does the emptiness is exposed.
The meaning you're going for here is that although there's a "bubble" it's likely a smaller bubble, not a large bubble like the dotcom bubble or the housing bubble.
The housing example is a great one. The 2008 housing bubble was due (in part) to overvalued houses. Those houses still have value though (they weren't scam houses that only existed on paper).
I never said the bubble implies the underlying asset has no value. The bubble encases the difference in value and current value. The bubble does not encase the actual asset.
When a bubble pops that difference disappears. But the asset value remains because the intrinsic value was not what the bubble was referring to.
I think he's probably right that we're very early on the adoption curve.
(even though I always get downvote and call a dumbass when I mention it here, I still think progress in quantum computing (predictive) and AI go hand in hand towards AGI.)
Basically every AI ( so to speak ) before ChatGPT is like Smartphone 1.0. Blackberry, Sony Ericsson, Nokia Symbian etc. ChatGPT is the Smartphone 2.0, aka iPhone era.
It look iPhone many years to reach mass adoption. MKBHD, as tech savvy as he is only had his first Smartphone in iPhone 4S era. ( If I remember correctly ). And then Phablet ( Any Screen larger than 5" ) came. I think we are at the end of that curve, 15 years after the iPhone introduction.
I think ChatGPT will follow a similar path. And if it was iPhone that pushed TSMC to sustain our current semiconductor improvement for the past ~10 years, then it will be ChatGPT that pushes us towards 1A ( 1nm ) node for another decade.
[1]https://wccftech.com/nvidia-ceo-calls-chatgpt-as-one-of-the-...
You get downvoted because it's not a coherent position. It's like saying I'm really bullish on renewable energy powered by web3 blockchains. It's just a mishmash of buzzwords. The reality is Quantum Computing has limited use cases in general computing and AI isn't really one of them. I don't think there's even an argument that e.g. a quantum chip offers any real advantage over a TPU or even GPU for AI tasks.
But being charitable I suppose if you subscribe to the idea that creativity and consciousness arise from quantum behaviours then perhaps it makes sense. It would suggest we likely have the architecture of an AGI completely wrong currently though, so I'm not sure how this really relates to AIs like ChatGPT.
Certainly a kind of god of the gaps reasoning, but actually quite relevant to our discussions of ai. If the world is fully deterministic, there's little room for the independent thought and free will we thinks makes our human existence special. We are then maybe just relatively sophisticated prediction engines, and even an LLM like ChatGPT isn't fundamentally dissimilar to us.
But "Quantum Computing AGI" is a complete non sequitur. It's just not a thing that needs to exist, as I already said quantum computing doesn't bring anything new to that table for current AI technology(Deep Learning).
Maybe we need AI to figure out quantum computing.
MS Teams gets the GPT treatment. It watches all your chats, email, calendars, code, wiki, meetings, spreadsheets, documents. Where is Slack now?
If you want to know absolutely anything from engineering domain knowledge to product strategy through to aiding you in sales, Slack has no answer. Ask Teams to splat you a database table, refine an SQL statement, brief you on a meeting, remind me of who a particular customer is and what sort of sales pitch would appeal to them.
I mean, not seeing the potential in GPT is really being intentionally blind to world changing technology. Just the fact that it can scan your whole codebase, find potential security holes, suggest performance blind-spots and indeed write code, or at least suggest code, this alone is such a big change it's hard to get your head around the opportunity. All of that in a chat window or IDE. It's revolutionary.
At that point instead of your boss asking you to send an email to someone or asking the data team to pull some stat, they'll just ask the chatbot to do it for them.
My guess is 80%+ of the work most people in corporate jobs do could fairly easily be automated with the next generation of GPT being fully integrated with an organisations data and tools.
It's so obvious the power of this technology. Those saying ChatGPT is still making a few programming errors with their crappy prompts are missing the point. Wait until a slightly more advanced version of GPT has access to your dev documentation + all your repos + your jira ticket board + your dev environment.
You won't even need to ask it to do anything. Your boss is going to quickly wonder why they need a team of 20 devs when a team of 2 devs reviewing ChatGPT pull requests is 10x more efficient.
Color me jaded, but I don’t care if the crud feature du jour is full of tech debt when we’re gonna throw it out in 3 months anyway.
Yes, and therein lies the issue - that last crucial 1%, 0.1% may be impossible to achieve (sort of like attaining lightspeed travel)
The fact some people argue ChatGPT already is already good enough for a lot of use cases should indicate we're not that far from something just as reliable as a human programmer.
I predict the Internet will soon go spectacularly supernova and in 1996 catastrophically collapse.
- Robert Metcalfe, in InfoWorld, 1995
Most things that succeed don’t require retraining 250 million people.
- Brian Carpenter, in the Associated Press, 1995
Tim Berners-Lee forgot to make an expiry date compulsory . . . any information can just be left and forgotten. It could stay on the network until it is five years out of date.
- Brian Carpenter, in the Associated Press, 1995
why you are so confident chatgpt ever will be able to work as independent dev and won't hit the limit of it's abilities?
Almost all tech has room for progress, ChatGPT is no different, it's just a question of how fast that progress will be.
You can roughly predict the rate of future progress by looking at how quickly advancements have been made in the recent past. In my opinion AI looks like computers in the 70s/80s or cars in the 40s. The tech isn't completely new, but significant improvements are being made every year.
I think the burden of proof would be on you to explain why advancements in AI would stop right here. Conveniently right around the time when AI is able to write code fairly well but with a few bugs.
all tech have also its limits, we could use internal combustion engines in the car, and even in some airplanes, but we have problems with this tech in building fighter jets and then spaceships.
The same is about ChatGPT, it is good to spill texts it seen during the training, or do slight transformation on it, but there were multiple attempts to teach NN do some algorithmic work, and it always failed miserably afaik.
lol, of course not each tech. You picked cars as example, but it was one of the most transformative advancements, and for each such advancements there are hundred thousands with moderate impact, and tens millions failures.
also, it is yet to be proven that ChatGPT will do well in any job requiring significant reasoning skills.
I'd have thought that people on HN would have far more vision than this. For every single "well it can't do that yet" observation there are dozens of use cases people are finding TODAY. It's already widely useful and this is year one.
discussion started with person claiming that soon chatgpt will be trained on docs, code and jira tickets and completely replace engineers, which imo will require significant reasoning skill.
I agree that for those tasks you described, chatgpt may find its niche.
That was actually me. And it already can write Jiras, scaffold code, and help write docs. Doesn't need any significant reasoning, just a well trained model and contextual data.
However, I didn't say it was going to 100% replace anyone. The horizon looks like GPT will become a significant assistant tool for numerous tasks. I think it will need humans to review and approve output and direct it for the foreseeable future.
yes, and my hypothesis that's where it will stop, because core meaningful engineering work on complex product/system requires way more reasoning abilities.
That's like saying junior developers will never take your job because they just don't know enough context or have enough experience.
Why do you think there is any reason for progress to stop? When has progress ever stopped? Even just looking at the simplistic Github Copilot from ~3 years ago I could see the writing on the wall for jnr devs. When these models have your entire codebase consumed it's quite apparent that the gestalt has changed in a way I think you are underestimating.
Complex product/systems is exactly where you'll need an AI code writer. It knows all of everything in your massive codebase. It will be able to suggest efficiencies you couldn't possibly be aware of unless you'd read every line of code yourself.
Remember when AlphaGo made that one move that showed it was far, far superior than a human at improvising and seeing ahead leagues of moves? It shocked the whole community, even the DeepMind engineers were shocked. That's going to be you one day. There is absolutely no reason I can see for this not to happen bar some sort of as yet undiscovered logical limit.
I described you reasons: jun developer has proven abilities to learn reasoning and context. and chat gpt does not have proven abilities to learn to reason.
All well and good when it's usually your enemy's space on the chopping block but mainstream media haven't yet told people if this kind of progress is generally desirable. The flock is flying blind so they default back to conservatism... for now.
I am having a hard time seeing this.
Intranet/corporate search has forever been awful in comparison to internet search.
-----------
> Employee: Hey CorpGod; where is that document that had some information about our new project management processes? You know, the one I had open like 2 weeks ago
> CorpGod: Oh, you mean the one your boss asked you to read by Friday? It's here [link]
> Employee: No, that's not it. It's the one that we copied and made edits to after that, I can't find it.
> CorpGod: That got deleted after Employee2 made a copy and published it. It's here on the Wiki [link]
-----------
GPT already does stuff like that with data from the Internet.
1) Scaling 2) Efficiency 4) Quality 5) Security
Some of these may be super-duper hard problems, but hard problems worth solving = massive opportunity.
If IBM could put space age technology into corporate offices in the 50s, we can put dystopian era technology into offices in the 2020s. At first it will be stupidly expensive, and only the big players will be able to take advantage of the cost to benefit ratio, but in time it will more than likely be on your phone.
Except that we are just left with outputs that are untrustworthy. All of these GPT products; ChatGPT, Copilot, Bard, Bing AI are still frequently hallucinating answers and often very incorrect solutions. We have already seen this with Copilot writing vulnerable code.
What this current AI hype cycle fails to realize is that given that you still cannot trust the generated output, it cannot be used safely in serious and highly regulated industries such as finance, law and medical professions all of which require trust and have been subject to AI disruption for years all with the same problem of trust being unsolved in AI. It is not enough to even disrupt search engines.
There is nothing new or revolutionary about a AI SaaS business with an API with a chatbot generating nonsense. I expect the hype around AI LLM chatbots to subside just like the hype around social spaces apps like Clubhouse did.
Trust in information is for people who outsource their every opinion, all they want to know is if it will keep them in high esteem for re-stating it and blinding following it. Well, for law, health, finance, that depends what year you got your opinion since the best information changes. Which is what we want if we want better.
A.I. output is just Words on a screen, it only promises coherence. How well a technology assists you is up to you or else we'd call it a torture device.
Seems like this is an interesting engineering or product problem worth solving. History is littered with big problems that were solved. Go look at flight, within 50 years we went from "it's not possible", to it's too fragile, to jets, to international airports and mass transit never before possible.
If you work in tech I think you've slept through your life.
Trust is a social problem, Not an engineering problem. Without a fundamental breakthrough in neural networks in transparency, the use-case for LLMs as search engines will always be eternally untrustworthy.
Not even the Bing AI that Microsoft released is even a trustworthy search engine [0] . In fact it is less trustworthy than Google.
There will be other, interesting use cases for sure, but control, compliance, documentation, etc. will be big.
It's cynical because you are viewing everything through one personal perspective that is loosely grounded in reality, and choosing to not see the rest of the world, which only serves to validate your cynical mindset.
For example, modern controlling functions couldn't exist without a certain level of technology. How much technology spend these days is dedicated to some form of administrative tasks? Those $$$ will have an effect.
In the real world, you can’t have a model that slurps up every bit of information in a company and then just lets anybody ask open ended questions about it.
But the security solutions for these technologies are far from maturity. They’re almost certainly addressable, but its going to be a whole industry in itself, will take years to take shape, and will probably involve underlying architectures that are designed very differently from what we see in this generation of models.
These corporate AI instances will take huge configuration, and like when IBM started adding infrastructure to large corps it was a huge effort. But the reward and advantage it provided was worth the millions of dollars and years of work. Eventually it all permeated down to consumers.
How long this will take it up for debate, but I don't think we can easily dismiss that it's inevitable.
It may not be what you describe but they are definitely flirting.
AI is only as good as the data it's learned from. There is little to no value in a really super awesome prompt. "Prompt engineering" is not a real thing. There is, however, enormous value in an AI that has learned on some specific set of data that nobody else has access to.
IOW, data is still the currency.
Correct. Everyone and their cats are now an AI company again. Hyping and parading about a hallucinating chatbot is going to change the world and take over search engines and kill Google. It won't. It needs to do more than that to even challenge Google.
We are already looking at its limitations and after looking at both Bing AI, ChatGPT and Google Bard, they all fall short at reliability and it is all fundamentally rooted to the black-box nature of neural networks.
The hype and mania will go on just as long as how Clubhouse was hyped on for in 2020 and like what happened to GPT-3 after that AI hype cycle died.
As the technology progresses and models are optimized/shrunk, I'm not sure if these "AI companies" will ever stand a chance. Even cheap Android smartphones can run the smallest GPT-Neo model, eventually the need for the SAAS wrappers for the technology will be cannibalized.
However as usual people will get too excited and the hype will outpace the actuality of the technology. There will be a slight "bubble" but this isn't anything like the "housing bubble".
Prior to chatGPT though, ai could be characterized as something along the lines of a housing bubble. I would say almost all lines of research in ai save llms are over hyped bubbles.
Not saying these lines of research are useless or inconsequential. Far from it. Ai outside of llms is amazing. But these Ais are definitely inside huge bubbles.
Any such language would necessarily be limited to their 'domain of existence' - you can't invent words for colours if your world has no light. Thus we'd need to give the AIs a full domain of existence for a full AI. They would need eyes (and ears?) and locomotion so they have a real world to reason about and talk about to each other (and us?)
The point being that emergent language is the only general way to gauge intelligence (above and beyond the somewhat anachronistic Turing test). I also conjecture that human language (in fact any human language) is complete in the sense that any and everything can be described within it (e.g. you could explain General Relativity to any human from any time, using their own language, as the basic concepts are already present, you just build on them - whereas this might not be possible with bird calls).
Furthermore, if our language is indeed complete, then we could suggest that any intelligent alien species we might encounter will also have at best a similarly complete language - and thus cannot be meaningfully more intelligent than us, as there is no 'higher' language, no concepts inaccessible to us.
For something to be overhyped, it must have a hype level disproportionate to its value. Deep learning is generating a huge amount of real value and is solving real problems. When we were getting Superbowl ads for crypto, the value proposition wasn't there.
Sometimes it was fraud, other times it was just money-losing economics ("traction") where VCs would prop it up and then hoped to dump the stock on an eager public in an IPO before making a dollar in profit (SoftBank took this to the extreme). Other times it just led to monopolies that sucked up everyone's data, engaged in surveillance capitalism, distracted them at dinner and made everyone have the attention of a goldfish. Oh yeah and made them depressed and insecure, especially teenagers.
I'm not sure that's much better than Web3.
Don't get me wrong, there definitely is excitement and hype, but it's all off the back of what's available today. There is no bubble to pop. If ChatGPT and diffusion models are the upper limit of the tech, it can still be applied today to disrupt whole industries.
That's not too say there won't be waves of grifters —some of them AI— fluffing products that never materialise or under-deliver, but I think our imagination is going to lag behind the tech, not the other way around.
FTA:
"Walmsley adds that he’s heard venture-capital investors speculate that the market for generative AI applications could be as large as $1 trillion. He notes that the world has over one billion knowledge workers; OpenAI charges $42 a month for the professional version of ChatGPT. If you assume every one of those people gets two accounts—one general, and one specialized—you get close $1 trillion."
All "AI" ETF that i see are filled with generic IT companies...
Anyone found a good one yet ?
So what happens is that awful companies that happen to be already public will try to reinvent themselves as part of the trend. Some past examples:
- Fish-oil company Zapata happened to own the domain zap.com, so in 1999 they became an Internet portal: https://www.forbes.com/1999/04/13/mu11.html?sh=5d90ad3b65d8
- In 2017 a beverage company called Long Island Iced Tea changed its name to Long Blockchain Corp: https://www.cnbc.com/2017/12/21/long-island-iced-tea-micro-c...
With mostly companies of this quality to pick from, anybody who built an Internet portfolio in 1999 or a blockchain portfolio in 2017 probably lost 99% of their money.
BitCoin had no real use-cases beyond betting and pump and dump and niche trading that the general public could not get in on. ChatGPT is being used creatively today by millions. It does your homework.
How is this classed as a bubble or hype. How can you possibly see it as a waste of effort or equivalent to talking to a parrot??
I can't use it to generate anything without thorough review for soundness either, so it doesn't really have potential to automate much of anything as far as I'm concerned.
About the only really good use of it I have seen is "Nothing, Forever", an entertainment/art piece that thrives on the jank of what it produces.
I have no doubt that technology will eventually get there, but I certainly don't see it as there today.
"Yeah, I hear what you're saying about the limitations of language models like GPT-3, especially when it comes to more creative tasks like writing. But it's still pretty impressive what these models can do! I mean, think about all the time they can save us by handling repetitive tasks, like customer service or even generating news articles. And who knows, maybe someday they'll be even better at the creative stuff too.
I know it's frustrating when you have to review everything the model outputs, but that's not just a problem with language models. People make mistakes too, y'know? So let's use these models as tools to help us, instead of trying to replace us completely.
All in all, while they may not be perfect yet, language models have a ton of potential to make our lives easier and more efficient. And that's pretty cool, if you ask me!"
This is probably helpful to someone, presumably someone with a great need for an automated bullshit generator, but it isn't helpful to me.
> I mean, think about all the time they can save us by handling repetitive tasks, like customer service or even generating news articles.
Something robots already do and people already hate them for it[0]. I doubt anyone will empower the robot to do anything meaningful, so while it can shape a word salad with a high degree of sounding like a human being, it only sounds like a useless human being (for the thing I care about). I'd be better off looking at whatever data it has that it is supposed to be using to help me.
> People make mistakes too, y'know? So let's use these models as tools to help us, instead of trying to replace us completely.
I'm perfectly capable of making mistakes on my own, thanks. And as anyone who's ever had to train anyone can tell you, looking over someone else's work constantly doesn't exactly help you get anything done. Give me a GPT I can train myself to eventual competency and maybe it'll pay dividends, but that doesn't realistically exist today.
[0] And of course another application is to use another GPT instance to summarize the bullshit news articles that were themselves generated from a summary, or interact with the usless customer service chatbot, thereby closing the cycle of waste that has become the cornerstone of modern civilization.
This is demonstrably not true. It's saying quite a bit, yes it's rather PR-speak like, but that's how it's been trained.
> And as anyone who's ever had to train anyone can tell you, looking over someone else's work constantly doesn't exactly help you get anything done
WTF? I think the way you are thinking about this seems fundamentally bogged down by cynicism. Proof reading an article written for you is far, far speedier than writing it yourself. That's why newspaper editors a) edit, b) get someone to write for them to edit. I can review a PR far more quickly than writing the code.
> I'm perfectly capable of making mistakes on my own, thanks
But now I don't need to pay you to make mistakes, GPT will make them instead and whoever was checking on your mistakes is still useful.
> Give me a GPT I can train myself to eventual competency and maybe it'll pay dividends, but that doesn't realistically exist today.
And GPT didn't exist in the wild last year. You have just admitted this is WIP, and I see no reason not to expect progress will continue. Maybe one day we'll all have personally trained AI models that live through our whole lives with us. They'll be able to give personal life advice, around to bounce ideas off of, help you remember everything. I mean it's trivial to find revolutionary use cases for this tech. The only question is "is it possible or is there a hard wall in front of us?" I am yet to be convinced there is anything standing in the way and billions of dollars is backing that up right now.
Honestly GPT here has more insight than you do, I think that's miraculous.
It has used many words, it has said very little, none of it useful or insightful. If you see something insightful in there, I must chalk it up to very different perception of reality.
> WTF? I think the way you are thinking about this seems fundamentally bogged down by cynicism. Proof reading an article written for you is far, far speedier than writing it yourself.
Not if you keep screwing it up. Not if I can't ever trust that you even tried to source correct information. Not if your creative output is trite at best.
> You have just admitted this is WIP, and I see no reason not to expect progress will continue.
I clearly said several times that I expect the technology will get better.
This isn't a hard problem (it's a few dozen lines of code to do the bare minimum, maybe a couple of hundreds to add a reasonably full featured parser), but it involved a conversation of refining it that wasn't just "fill in this missing detail", but e.g. asking it to implement parsing of the query package in a specific language (Ruby), using a specific API, and using a specific format. Ruby ships with a DNS implementation; the ChatGPT solution looks nothing like either the standard Ruby implementation or the alternative Ruby DNS implementations I've checked.
You must know parrots with better reasoning skills than I do.
[EDIT: Note that I specifically picked this exactly because I was sceptical of how well it'd do, it's something I've done before myself, with a protocol I know well enough to be able to evaluate the code without much effort]
Not, but a tree can.
> emulate theory of mind,
Neither can Chat GPT. "Failing a Theory of Mind at the same level a 7 year old does" is no more Theory of Mind than being really good at chess is.
> or implement merge sort with pirate themed variable names.
True.
Now we are hitting an inflection point. People are realizing chatGPT is more then that.
The people getting voted down are the people saying chatGPT is just some trivial word generator. The masses are realizing that chatGPT is different.
There's just been years and years of AI buzz so people are used to downplaying it. So much so that when something genuinely intelligent is built they just fall into the same generic pattern of saying that these things are just statistical word generators because "oh it got my specific question wrong." It's biased because they are dismissing everything it gets right.